Automating Industrial Quality Assurance: A Zero-Shot MLLM Framework for Defect Detection and 3R

In modern industrial production, particularly within the circular economy, automated defect detection is crucial for quality control and resource efficiency. Traditional inspection methods are often inconsistent and inefficient, while existing machine learning approaches are constrained by data scarcity, poor generalization, and a lack of interpretability. This paper proposes a zero-training visual inspection framework that leverages Multi-Modal Large Language Models (MLLMs) for automated defect detection and decision-making in industrial components, using starter motors as a case study. The framework integrates the zero-shot reasoning capabilities of an MLLM with an Elasticsearch-powered knowledge base to create an end-to-end automated system for "defect identification, standard matching, and decision-making." By processing images and relevant technical documentation, the system identifies defects and provides handling recommendations based on the 3R principles (Reuse, Remanufacture, Recycle) without requiring task-specific training. Experimental results on 584 starter motor samples show that our optimized system achieves an overall accuracy of 81.4% and an F1 score of 85.3%. This work validates the potential of MLLMs for industrial visual inspection and provides a replicable framework for advancing intelligent automation in manufacturing.